Knowledge Information Flashcards

What is the primary difference between 'parametric knowledge' and 'non-parametric knowledge' in AI systems?

Answer: Parametric knowledge is stored in model weights; non-parametric knowledge is stored in external databases

Parametric knowledge is encoded in a model's learned weights during training, while non-parametric knowledge is stored externally (e.g., in a vector DB or document store) and retrieved at inference time.

Which SPARQL query keyword is used to retrieve triples from an RDF knowledge graph?

Answer: SELECT with WHERE patterns

SPARQL uses SELECT queries with WHERE clauses containing triple patterns (subject, predicate, object) to match and retrieve data from RDF graphs.

What is 'multi-hop reasoning' in knowledge graphs and QA systems?

Answer: Answering a question by chaining multiple inference steps across graph edges

Multi-hop reasoning requires traversing multiple relationships in a knowledge graph or making several logical inferences to arrive at an answer that cannot be found in a single step.

Why might an AI engineer choose hybrid retrieval (sparse + dense) over pure dense retrieval in a production RAG system?

Answer: Hybrid retrieval combines lexical exactness with semantic understanding, improving coverage

Hybrid retrieval leverages BM25's strength in exact keyword matching alongside dense embeddings' semantic similarity, reducing failure cases where one method alone would miss relevant documents.

What is the role of 'reranking' in a two-stage retrieval pipeline?

Answer: Re-scoring a small candidate set using a higher-capacity model to improve final ranking quality

A reranker (e.g., a cross-encoder) takes the top-k candidates from a fast first-stage retriever and applies a more computationally expensive relevance model to produce a higher-quality final ranking.

Which approach is most effective for detecting and reducing 'factual drift' in long-context language model outputs?

Answer: Applying self-consistency checks or chain-of-thought verification against source documents

Self-consistency methods and chain-of-thought prompting that explicitly cites source passages help detect when generated claims diverge from grounded facts as output length grows.

In the context of knowledge management for AI systems, what does 'provenance tracking' enable?

Answer: Tracing every fact back to its original source for auditability and trust assessment

Provenance tracking records the origin, lineage, and transformation history of each piece of knowledge, enabling auditors and systems to assess source reliability and update or retract facts when sources change.